{"id":"01ddf947-26af-41ad-812e-ab51264656c4","arxiv_id":"2508.02079","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":3,"one_line_summary":"AlignGuard-LoRA is a LoRA variant that uses Fisher information and collision-aware geometry to reduce alignment drift while preserving downstream task accuracy.","lead":"AlignGuard-LoRA adds alignment-preserving penalties to LoRA fine-tuning, and the authors report up to 50% less safety drift on their DriftCaps benchmark. The paper also claims a scaling law for catastrophic forgetting, but the body text could not be read in this submission.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Fisher-at-init may not identify alignment-sensitive directions; without a random-subspace control, 'up to 50% drift reduction' does not establish that AGL's Fisher/geodesic mechanism is doing the work.","rationale":"The reader's verdict is UNVERDICTED because only the abstract was machine-readable. My read of the abstract and the corrupted full text finds no internal inconsistency that would justify outright rejection, but it does expose one load-bearing unverified condition: Fisher information computed at base-model initialization must actually identify the parameter directions that matter for alignment. This is not a trivial assumption. The FIM's top eigenvectors describe local parameter sensitivity at one point, while alignment is an input-dependent behavioral property that may live in directions with small Fisher information. If the method merely dampens updates, then a random low-rank projection of the same rank could reproduce the claimed drift reduction, making the Fisher/geodesic machinery decorative rather than causal. The abstract's ablation statement that each component 'contributes distinctly' does not establish the mechanism unless a random-projection placebo is included. I therefore propose one decisive computational check using the authors' own open-source implementation. Because the full text cannot currently be read, derivations cannot be checked and no accept/reject decision is warranted; the responsible verdict remains unverdictable. I do not allege any issue with the authors' integrity, and I credit the paper for releasing code, which makes the proposed test feasible.","tokens_in":8081,"tokens_out":4813,"duration_ms":63233,"concrete_test":"Run a placebo substitution in the released implementation: replace the Fisher-based projection with a fixed random orthogonal projection of the same rank while preserving all other terms, and match the total LoRA update norm across conditions; evaluate on DriftCaps over 10+ seeds. If random-projection AGL gives statistically indistinguishable drift reduction, the Fisher/geodesic mechanism is not doing the work; if it is clearly worse, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that the base model's Fisher Information Matrix at initialization marks the parameter directions whose perturbation causes alignment drift. That premise is not demonstrated and is vulnerable in a specific way: high-Fisher directions are local curvature directions, not necessarily safety-behavior directions. Safety constraints can be enforced by low-curvature or input-dependent mechanisms, and the FIM is estimated at a single point, so its top eigenspace can be noisy and unstable. If AGL's main effect is simply to keep the LoRA update small, or to restrict it along any low-dimensional subspace, then the reported 50% drift mitigation would occur without the Fisher-guided or Riemannian geometry contributing anything. The ablations claimed in the abstract cannot rule this out unless they include a placebo projection baseline. Moreover, the supplied full text is not machine-readable, so neither the derivation connecting Fisher eigenvectors to alignment sensitivity nor the claimed scaling law can be checked; the paper is currently unverdictable.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes AlignGuard-LoRA (AGL), an extension of LoRA for fine-tuning large language models that aims to preserve alignment and safety behavior. AGL adds Fisher Information Matrix-based regularization to restrict updates in alignment-sensitive subspaces, task-specific regularization, and a collision-aware regularization combining Riemannian overlap and geodesic separation. The authors also introduce DriftCaps, a diagnostic benchmark of safe and unsafe prompts, and claim empirical results that AGL mitigates alignment drift by up to 50% without degrading downstream performance. In addition, the abstract states that a scaling law for catastrophic forgetting is derived and validated. The supplied full text, however, is heavily corrupted and not machine-readable, so the technical details and experimental results cannot be verified.","tokens_in":8294,"tokens_out":2746,"duration_ms":34846,"significance":"If the central claims hold, AGL would be a practical, drop-in refinement to LoRA that reduces alignment drift during fine-tuning, and the proposed scaling law for catastrophic forgetting could be a useful tool for predicting loss escalation. The paper also contributes a new benchmark (DriftCaps) and an open-source implementation, both of which could benefit the community. However, the current manuscript cannot support these claims in verified form: the body of the paper is unreadable, the headline improvement is reported without error bars or baseline details, and the methodological premise that the Fisher Information Matrix at initialization identifies alignment-sensitive directions is not tested against a placebo control. The significance is therefore conditional and cannot be established from the submitted text.","major_comments":[{"comment":"The supplied full text is not readable: most of the body is replaced by mojibake and line noise, so the derivations of the Fisher-guided decomposition, the Riemannian-geodesic collision regularizer, the ablation methodology, and the scaling-law fit cannot be checked. This is a blocking issue: the manuscript must be resubmitted with a readable text before the technical content can be evaluated.","section":"Full text (corrupted)"},{"comment":"The headline claim in the abstract, that AGL 'mitigates alignment drift by up to 50% on safety-critical benchmarks without degrading downstream task performance,' is reported without error bars, number of seeds, model sizes, or a baseline table. As written, this is a claim-without-visible-evidence situation; the empirical section cannot be verified because the full text is corrupted. The authors should provide a concrete table with means and variances across seeds, and specify the baselines (e.g., standard LoRA, LoRA with weight decay, and a random-subspace regularizer).","section":"Abstract"},{"comment":"The premise that the Fisher Information Matrix computed at initialization identifies alignment-sensitive parameter directions is not demonstrated. High-Fisher directions reflect local curvature, not necessarily safety-relevant behavior; safety constraints may be encoded in low-curvature or input-dependent ways. Without a placebo-projection baseline that restricts updates to a random low-dimensional subspace matched for update norm, the claimed 'up to 50% drift reduction' does not establish that the Fisher guidance or the Riemannian geometry is doing the work. The ablation described in the abstract cannot rule this out unless such a control is included.","section":"Method (Fisher regularization)"},{"comment":"The abstract states that a scaling law for catastrophic forgetting is 'derived and validated' using the same author-curated DriftCaps benchmark on which the method is also evaluated. This creates a circularity risk: if the scaling-law parameters are fitted on the same data used to demonstrate the effect, the validation is not independent. The paper should either derive the law from a separate dataset or explicitly validate on a held-out set or on an unrelated benchmark.","section":"Scaling law (as described in abstract)"}],"minor_comments":[{"comment":"The phrase 'up to 50%' is vague; the paper should state the specific metrics, benchmark names, and conditions under which this maximum is achieved.","section":"Abstract"},{"comment":"The terms 'Riemannian overlap' and 'geodesic separation' are introduced without definitions in the abstract; a brief intuitive explanation would help readers understand the proposed mechanism.","section":"Abstract"},{"comment":"The benchmark 'DriftCaps' is mentioned but not described even at a high level (e.g., number of prompts, domain coverage, or how alignment drift is quantified); a short description would improve clarity.","section":"Abstract"}],"recommendation":"uncertain","confidential_remarks":"The provided full text is so corrupted that no substantive technical review is possible. This appears to be an encoding/extraction artifact rather than a scientific failure by the authors, but as a matter of process the manuscript is unverdictable. I recommend requesting a clean version before any further editorial decision. Even in the abstract, the empirical claims lack the statistical and baseline detail needed to assess them; the absence of a random-subspace control is a substantive concern that should be addressed in revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The one thing to know: I can't actually read the paper. The copy I have is corrupted—garbled characters from top to bottom, with only the abstract and a few equation fragments intact. So this is a review of the abstract, not the work.\n\nWhat's new: the collision-aware geodesic penalty and the DriftCaps benchmark appear to be original. Putting a Fisher-based regularizer on LoRA is not new, but the Riemannian overlap term that penalizes coordinate-wise interference is a different tack. The authors also claim a derived and validated scaling law for catastrophic forgetting. Those are concrete contributions worth checking, and they say they've open-sourced the implementation.\n\nWhat I can't verify: the headline \"up to 50% drift reduction\" comes with no error bars, no model scale, no baseline table in the abstract. The stress-test worry is on point—Fisher at initialization is a local curvature measure, and the directions that matter for safety could be low-curvature or input-dependent. The ablations, as described, won't settle this unless they include a placebo projection baseline (e.g., a random subspace of the same dimensionality) and a simple weight-norm control. If reducing the LoRA update norm alone gets most of the benefit, the Fisher/geodesic machinery is decoration. I also note the scaling law is derived and validated on the same setup, which risks circularity unless they test it on a held-out model or task.\n\nNone of this is an indictment—these are questions any competent referee would ask. The paper is coherent at the abstract level and the claimed components are distinct. But the provided full text is unreadable, so I cannot speak to the math or the experiments.\n\nWho this is for: anyone working on safe fine-tuning or LoRA variants. If the real PDF is clean, it deserves a serious referee. My advice: engage, but condition review on a readable manuscript, and make sure the reviewers push for a random-subspace baseline and cross-model/cross-task error bars. Right now I wouldn't cite it and I wouldn't bet on it, but it's not a dismiss.","headline":"Plausible-sounding LoRA safety method with a new benchmark and scaling law, but the supplied full text is unreadable, so I'd conditionally send it to review and push for a random-subspace control.","tokens_in":8740,"tokens_out":2778,"would_cite":false,"duration_ms":34075,"reading_group":"maybe","serious_thinker":"unclear","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A Fisher-guided LoRA variant cuts alignment drift by up to 50 percent during fine-tuning.","keywords":["low-rank adaptation","alignment","fine-tuning","Fisher Information Matrix","Riemannian geometry","catastrophic forgetting","LLM safety","regularization"],"falsifier":"One decisive test would be to compute the overlap between the top eigenvectors of the base-model Fisher Information Matrix and the gradient of a safety-alignment loss on a model and task where vanilla LoRA shows large drift; if the overlap is near zero, AGL's Fisher regularizer cannot protect that safety objective, and any observed drift reduction would have to be explained by the other components.","tokens_in":7936,"feed_emoji":"🛡️","tokens_out":3390,"duration_ms":43688,"temperature":0.7,"pith_summary":"This paper tries to establish that alignment drift caused by LoRA fine-tuning is not inevitable: if parameter updates are restrained in directions the base model treats as alignment-sensitive, drift can drop by up to half on safety-critical benchmarks while downstream task performance holds. The method, AlignGuard-LoRA (AGL), adds three regularizers to the usual LoRA loss, with the load-bearing one built from the Fisher Information Matrix of the base model. The paper also introduces DriftCaps, a curated benchmark of safe and unsafe prompts, and derives a scaling law for catastrophic forgetting that predicts post-fine-tuning loss escalation. If the claims hold, AGL is a drop-in refinement of LoRA that protects safety behavior without sacrificing adaptation quality.","feed_headline":"Fisher-guided LoRA cuts alignment drift by up to 50%","feed_subtitle":"A new regularizer keeps safety behavior intact while LoRA adapts models to new tasks.","key_machinery":"The load-bearing object is the Fisher Information Matrix of the base model, computed before fine-tuning; it encodes which parameter directions the model's output distribution is most sensitive to, and AGL uses it to keep LoRA updates out of alignment-critical directions. The second mechanism is collision-aware regularization, which combines a Riemannian-overlap term that penalizes coordinate-wise interference between new-task updates and existing behavior with a geodesic-separation term that pushes the update path into a distinct region of parameter space. Together these terms redirect the update rather than merely shrinking it, preserving adaptation while protecting safety behaviors.","core_discovery":"The central claim is that AlignGuard-LoRA (AGL) preserves latent safety behavior during LoRA fine-tuning by treating alignment as a set of sensitive parameter directions rather than as a separate objective. AGL augments the task loss with a Fisher Information Matrix-based regularizer, computed at initialization, that penalizes movement in alignment-sensitive subspaces; a task-specific regularizer that stabilizes the integration of new knowledge; and a collision-aware regularizer that blends Riemannian overlap, which penalizes coordinate-wise interference, with geodesic separation, which encourages updates to live in a different region of parameter geometry. On DriftCaps, the paper's diagnostic benchmark of safe and unsafe prompts, AGL reports up to 50 percent reduction in alignment drift without degrading downstream performance. Comprehensive ablations attribute the improvement to each component, and a derived, validated scaling law indicates that AGL flattens the loss escalation that normally follows fine-tuning.","pith_inferences":["The method's protection is only as good as the static Fisher Information Matrix computed at initialization; a stronger variant might update the Fisher estimate online during fine-tuning, an extension the paper does not test.","The geodesic-separation idea likely generalizes beyond safety, for instance to continual learning, where new tasks must avoid overwriting old task geometry in any multi-objective setting.","DriftCaps could be used to test whether the reported drift reduction extends to out-of-distribution adversarial prompts, a regime beyond the benchmarks reported in the paper."],"forward_implications":["Safety-aligned LLMs could be fine-tuned with LoRA without post-hoc safety retraining or additional alignment data.","DriftCaps provides a standardized diagnostic for quantifying alignment drift, enabling apples-to-apples comparisons across fine-tuning methods.","The derived scaling law for catastrophic forgetting lets practitioners predict when post-fine-tuning loss escalation sets in and choose a regularization budget accordingly.","Each of AGL's regularizers can be adopted separately inside existing LoRA pipelines, so parts of the method could be used even where the full framework is not."],"supporting_citations":[],"fun_headline_variants":["Alignment-safe LoRA: Fisher guard halves drift","Fisher-geometry LoRA: 50% less alignment drift","Keep LoRA safe: Riemannian guard cuts drift 50%","LoRA drift cut 50% via Fisher-guided alignment guard"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument rests on the base model's Fisher Information Matrix, computed before fine-tuning, correctly identifying the parameter directions that actually govern safety behavior; if that identification is wrong, the regularizer restrains the wrong directions and the measured drift reduction must come from something else.","fun_headline_variants_meta":{"raw":{"variants":["Alignment-safe LoRA: Fisher guard halves drift","Fisher-geometry LoRA: 50% less alignment drift","Keep LoRA safe: Riemannian guard cuts drift 50%","LoRA drift cut 50% via Fisher-guided alignment guard"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000524,"raw_usage":{"total_tokens":2551,"prompt_tokens":984,"completion_tokens":1567,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":600,"completion_tokens_details":{"reasoning_tokens":1498}},"tokens_in":600,"tokens_out":1567,"duration_ms":12734,"temperature":1.0,"reasoning_tokens":1498,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T05:09:35.945886+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"One decisive test would be to compute the overlap between the top eigenvectors of the base-model Fisher Information Matrix and the gradient of a safety-alignment loss on a model and task where vanilla LoRA shows large drift; if the overlap is near zero, AGL's Fisher regularizer cannot protect that safety objective, and any observed drift reduction would have to be explained by the other components.","supporting_citations":[],"review_version":1}